Commits

Breadcrumbs

History for

onmain

User selector

All users

All time

Commit history

Commits on Jul 24, 2026

• [recipe] feat: Retain model output from tinker forward backward (#7109)

  Show description for 2b47a68
  

  wyettzeng

  and

  huaye.zeng

  authored
  2b47a68View commit details

  Copy full SHA for 2b47a68

  View code at this pointBrowse repository at this point

Commits on Jul 23, 2026

• [trainer] fix: isolate TaskRunner logs from root logger changes (#7133)

  Show description for d8acd86
  Begunner

  authored
  d8acd86View commit details

  Copy full SHA for d8acd86

  View code at this pointBrowse repository at this point

• [trainer] feat: V1 replay buffer eviction/refill handling for stale, DAPO-filtered, and failed rollout groups (#7082)

  Show description for c88e875
  Begunner

  authored
  c88e875View commit details

  Copy full SHA for c88e875

  View code at this pointBrowse repository at this point

• [rollout, hardware] fix: fix rollout replica.py for device abstraction (#7128)

  Show description for 11c94ad
  kahlun

  authored
  11c94adView commit details

  Copy full SHA for 11c94ad

  View code at this pointBrowse repository at this point

Commits on Jul 22, 2026

• [doc] chore: update Ascend 950 references and disable nightly test (#7111)

  Show description for a35908c
  

  wangdongleix

  and

  wangdonglei

  authored
  a35908cView commit details

  Copy full SHA for a35908c

  View code at this pointBrowse repository at this point

• [fsdp, perf] feat: defer gradient sync during accumulation (#7095)

  Show description for 2423f80
  

  zhangxin81

  and

  codex

  authored
  2423f80View commit details

  Copy full SHA for 2423f80

  View code at this pointBrowse repository at this point

Commits on Jul 21, 2026

• [perf] feat: re-support torch profiler and optimize output naming (#7099)

  Show description for a07d3e3
  ETOgaosion

  authored
  a07d3e3View commit details

  Copy full SHA for a07d3e3

  View code at this pointBrowse repository at this point

• [veomni, fsdp] fix: skip manual model.to(device) under FSDP2 CPU offload (#7103)

  Show description for 2c9cf59
  cben484

  authored
  2c9cf59View commit details

  Copy full SHA for 2c9cf59

  View code at this pointBrowse repository at this point

• [megatron] fix: Qwen3.5 LoRA & MTP support (with Megatron-Bridge) (#5599)

  Show description for a32d635
  HollowMan6

  authored
  a32d635View commit details

  Copy full SHA for a32d635

  View code at this pointBrowse repository at this point

Commits on Jul 20, 2026

• [model, fsdp] fix: support Qwen3.5 linear attention under Ulysses SP (#6660)

  Show description for 6a6242f
  

  sunnweiwei

  and

  wuxibin89

  authored
  6a6242fView commit details

  Copy full SHA for 6a6242f

  View code at this pointBrowse repository at this point

• [fully_async] feat: support dynamic resource scheduling (#6556)

  Show description for 298fdc7
  

  zpltys

  and

  zhangshuai122

  authored
  298fdc7View commit details

  Copy full SHA for 298fdc7

  View code at this pointBrowse repository at this point

• [doc] fix: repair broken Ascend tutorial links and shell filename (#7074)

  Show description for 2f835c5
  chengminhua

  authored
  2f835c5View commit details

  Copy full SHA for 2f835c5

  View code at this pointBrowse repository at this point

• [trainer, perf] fix: include standalone rollout GPUs in throughput denominator for separate async (#7049)

  Show description for 32bb33d
  mikequan0425

  authored
  32bb33dView commit details

  Copy full SHA for 32bb33d

  View code at this pointBrowse repository at this point

• [ckpt] fix: added cuda stream synchronization in NCCL broadcast wait for complete (#7083)

  Show description for 61f9699
  

  3 peopleauthored
  61f9699View commit details

  Copy full SHA for 61f9699

  View code at this pointBrowse repository at this point

Commits on Jul 17, 2026

• [rollout] fix: handle malformed Qwen3 XML tool calls (#7044)

  Show description for e003163
  

  zzzzzzzxh

  and

  codex

  authored
  e003163View commit details

  Copy full SHA for e003163

  View code at this pointBrowse repository at this point

• [vllm] fix: restore vllm patch (#7076)

  Show description for d718b8c
  tardis-key

  authored
  d718b8cView commit details

  Copy full SHA for d718b8c

  View code at this pointBrowse repository at this point

• [ckpt] feat: add save_lora_only checkpoint support (#7061)

  Show description for b5021d3
  Secbone

  authored
  b5021d3View commit details

  Copy full SHA for b5021d3

  View code at this pointBrowse repository at this point

Commits on Jul 16, 2026

• [trainer] fix: exclude the never-routed final response token from rollout MoE load-balance metrics (#7075)

  zsnoob

  authored
  74a718aView commit details

  Copy full SHA for 74a718a

  View code at this pointBrowse repository at this point

• [hardware] fix: fix build_npu_attn_mask oom (#7055)

  Show description for 45ab5b8
  

  RichardFido

  and

  r1chardf1d0

  authored
  45ab5b8View commit details

  Copy full SHA for 45ab5b8

  View code at this pointBrowse repository at this point

• [ckpt,rollout] feat: sharded delta weight sync over NCCL for disaggregated rollout (#6974)

  Show description for 903d90c
  

  3 peopleauthored
  903d90cView commit details

  Copy full SHA for 903d90c

  View code at this pointBrowse repository at this point

• [ci, trainer] fix: Fix ci AssertionError for the environment variable CUDA_DEVICE_MAX_CONNECTIONS (#7040)

  Show description for 748858b
  

  pengnuoheng

  and

  pengnuoheng

  authored
  748858bView commit details

  Copy full SHA for 748858b

  View code at this pointBrowse repository at this point

Commits on Jul 15, 2026

• [ci] chore: remove vllm_ascend patch and fix ci (#7042)

  Show description for 6f2aa24
  wucong25

  authored
  6f2aa24View commit details

  Copy full SHA for 6f2aa24

  View code at this pointBrowse repository at this point

• [trainer] fix: gen_batch_size falls back to train_batch_size (#7051)

  Show description for 6d0f114
  Begunner

  authored
  6d0f114View commit details

  Copy full SHA for 6d0f114

  View code at this pointBrowse repository at this point

• [fsdp] fix: sync merged LoRA weights before context exit (#7014)

  Show description for 6e3c39d
  

  rongkunxue

  and

  codex

  authored
  6e3c39dView commit details

  Copy full SHA for 6e3c39d

  View code at this pointBrowse repository at this point

• [tool] feat: verl add rl insight (#6680)

  Show description for 5b2bfe7
  mengchengTang

  authored
  5b2bfe7View commit details

  Copy full SHA for 5b2bfe7

  View code at this pointBrowse repository at this point

• [ci] test: update ascend ci docker tag (#7035)

  Show description for f183d8c
  

  3 peopleauthored
  f183d8cView commit details

  Copy full SHA for f183d8c

  View code at this pointBrowse repository at this point

Commits on Jul 14, 2026

• [trainer, ckpt] feat: support streaming dataloader and async trainer checkpoint recovery (#7037)

  Show description for 49ed5e1
  Begunner

  authored
  49ed5e1View commit details

  Copy full SHA for 49ed5e1

  View code at this pointBrowse repository at this point

• [rollout, vllm] fix: stop the policy from sampling vision placeholder tokens (#7038)

  Show description for 42adda3
  zhshj0110

  authored
  42adda3View commit details

  Copy full SHA for 42adda3

  View code at this pointBrowse repository at this point

• [megatron] feat: support DeepSeek V4 GRPO (#6473)

  Show description for cf8005f
  HollowMan6

  authored
  cf8005fView commit details

  Copy full SHA for cf8005f

  View code at this pointBrowse repository at this point

• [tool, rollout] feat: support parameter sync steps in Skip Manager (#7032)

  Show description for 8c4696e
  mikequan0425

  authored
  8c4696eView commit details

  Copy full SHA for 8c4696e

  View code at this pointBrowse repository at this point

Commits on Jul 13, 2026

• [rollout] fix: disable vLLM multimodal processor cache for vLLM < v0.22.0 (#6932)

  Show description for e5687fc
  lxb007981

  authored
  e5687fcView commit details

  Copy full SHA for e5687fc

  View code at this pointBrowse repository at this point

• [megatron] fix: return 3-tuple under calculate_per_token_loss to fix MoE aux/z-loss grad blowup at CP>1 (#6836)

  Show description for 04bac4e
  

  EricMarcus-ai

  and

  gemini-code-assist[bot]

  authored
  04bac4eView commit details

  Copy full SHA for 04bac4e

  View code at this pointBrowse repository at this point

Commits on Jul 11, 2026

• [doc] fix: update the readme about kimi-checkpoint-engine (#7018)

  Show description for 688544d
  beirong8kmiles

  authored
  688544dView commit details

  Copy full SHA for 688544d

  View code at this pointBrowse repository at this point

Commits on Jul 10, 2026

• [model] fix: fix vocab_weights device and shard handling for qwen3_vl.py (#6659)

  Show description for ad2e3c2
  

  ocp8

  and

  zhibaiwang

  authored
  ad2e3c2View commit details

  Copy full SHA for ad2e3c2

  View code at this pointBrowse repository at this point

• [fsdp] fix: skip the whole-shard staging round trip in FSDP2 weight export (#7005)

  Show description for b7ff88e
  ChangyiYang

  authored
  b7ff88eView commit details

  Copy full SHA for b7ff88e

  View code at this pointBrowse repository at this point